[PATCH] D79744: clang: Add address space to indirect abi info and use it for kernels

[Johannes Doerfert via Phabricator via cfe-commits]

jdoerfert added a comment.

In D79744#2040380 <https://reviews.llvm.org/D79744#2040380>, @rjmccall wrote:

> In D79744#2040348 <https://reviews.llvm.org/D79744#2040348>, @jdoerfert wrote:
>
> > Drive by, I haven't read the entire history.
> >
> > In D79744#2040208 <https://reviews.llvm.org/D79744#2040208>, @rjmccall wrote:
> >
> > > I don't understand why `noalias` is even a concern if the whole buffer passed to the kernel is read-only.  `noalias` is primarily about proving non-interference, but if you can tell IR that the buffer is read-only, that's a much more powerful statement.
> >
> >
> > The problem is that it is a "per-pointer" attribute and not "per-object". Given two argument pointers, where one is marked `readonly`, may still alias. Similarly, an access to a global, indirect accesses, ... can write the "readonly" memory.
>
>
> Oh, do we really not have a way to mark that memory is known to be truly immutable for a time?  That seems like a really bad limitation.  It should be doable with a custom alias analysis at least.


`noalias` + `readone` on an argument basically implies immutable for the function scope. I think we have invariant intrinsics that could do the trick as well, though I'm not too familiar with those. I was eventually hoping for paired/scoped `llvm.assumes` which would allow `noalias` + `readnone` again. Then there is `invariant` which can be placed on a load instruction. Finally, TBAA has a "constant memory" tag (or something like that), but again it is a per-access thing. That are all the in-tree ways I can think of right now.

Custom alias analysis can probably be used to some degree but except address spaces I'm unsure we have much that you can attach to a pointer and that "really sticks".

>>> Regardless, if you do need `noalias`, you should be able to emit the same IR that we'd emit if pointers to all the fields were assigned into `restrict` local variables and then only those variables were subsequently used.
>> 
>> We are still debating the local restrict pointer support. Once we merge that in it should be usable here.
> 
> I thought that was finished a few years ago; is it really not considered usable yet?  Or does "we" not just mean LLVM here?

Yes, I meant "we" = LLVM here. Maybe we talk about different things. I was referring to local restrict qualified variables, e.g., `double * __restrict Ptr = ...`. The proposal to not just ignore the restrict (see https://godbolt.org/z/jLzjR3) came last year, it was a big one and progress unfortunately stalled (partly my fault). Now we are just about to see a second push to get it done.
Is that what you meant too?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D79744/new/

https://reviews.llvm.org/D79744